6 Sep 2013

Dmitry Bogdanov defends his PhD thesis entitled "From music similarity to music recommendation: Computational approaches based on audio and metadata analysis" on Friday 6th of September 2013 at 11:00h in room 55.309.

The jury members of the defense are: Markus Schedl (Johannes Kepler University), Josep Lluis Arcos (IIIA-CSIC), Emilia Gómez (UPF).

Abstract: The amount of music available digitally is overwhelmingly increasing. Vast amounts of music are available for listeners, and require automatic organization and filtering. In this context, user modeling, which consists in customization and adaptation of systems to the user's specific needs, is a challenging fundamental problem. A number of music applications are grounded on user modeling to provide users a personalized experience. In the present work we focus on user modeling for music recommendation, and propose a preference elicitation technique in conjunction with different recommendation approaches. We develop algorithms for computational understanding and visualization of music preferences. Our approaches employ algorithms from the fields of signal processing, information retrieval, machine learning, and are grounded in cross-disciplinary research on user behavior and music. Firstly, we consider a number of preference elicitation strategies, and propose a user model starting from an explicit set of music tracks provided by this user as evidence of his/her preferences. The proposed strategy provides a noise-free representation of music preferences. Secondly, we study approaches to music similarity, working solely on audio content. We propose a novel semantic measure which benefits from automatically inferred high-level description of music. Moreover, we complement it with low-level timbral, temporal, and tonal information and propose a hybrid measure. The proposed measures show significant improvement, compared to common music similarity measures, in objective and subjective evaluations. Thirdly, we propose distance-based and probabilistic recommendation approaches working with explicitly given preference examples. Both content-based and metadata-based approaches are considered. The proposed methods employ semantic and hybrid similarity measures as well as they build semantic probabilistic model of music preference. Further filtering by metadata is proposed to improve results of purely content-based recommenders. Moreover, we propose a lightweight approach working exclusively on editorial metadata. Human evaluations show that our approaches are well-suited for music discovery in the long tail, and are competitive with metadata-based industrial systems. Fourthly, to provide insights on the nature of music preferences, we create regression models explaining music preferences of our participants and demonstrate important predictors of their preference from both acoustical and semantic perspectives. The obtained results correlate with existing research on music cognition. Finally, we demonstrate a preference visualization approach which allows to enhance user experience in recommender systems.